Mandip Goswami

Acoustics meets audio machine learning

I sit at the intersection of industrial acoustics and audio machine learning — a combination that is rare and increasingly consequential as robotic environments, autonomous vehicles, and AI-native products all demand acoustic intelligence. At Amazon Fulfillment Technologies & Robotics, I lead acoustic safety infrastructure for one of the world's largest and most complex built environments — hundreds of fulfillment centers, thousands of robots, millions of square feet of noise. My work spans psychoacoustic modeling, NVH prediction at scale, room acoustics simulation, and applied deep learning for multi-source audio classification. I also contribute to the open research ecosystem through publicly available datasets, models, and benchmarks on Hugging Face — focused on room acoustics, machine listening, and spatial audio, and used by researchers across speech, audio ML, and acoustic simulation. My broader goal is to help build the next generation of AI systems that understand physical sound environments: robotics, AR/VR spatial audio, smart infrastructure, industrial monitoring, and acoustic digital twins.

Research used by the community, at scale

Open datasets with real-world reach. Every corpus ships with validation tooling, checksums, and reference baselines — and the community has adopted them at scale.

Papers & Preprints

Peer-oriented research on room acoustics, reverberant speech, and robust ASR — each paper paired with an open dataset, benchmark, or tool for full reproducibility.

Datasets

Download counts, likes, and update dates below are pulled live from the Hugging Face API every time this page loads.

Models

Fine-tuned models for robust speech recognition under real room acoustics.

What the acoustics look like

Real figures from the published papers — room-acoustics statistics, reverberant-speech spectrograms, and dataset coverage plots. Each is linked to its source preprint.

RIR-Mega — Validation Summary

Dataset validation across 50,000 simulated RIRs: RT60, DRR, and room-volume distributions, room-dimension histograms, and a Schroeder energy-decay curve (RT60 ≈ 0.26 s).

arXiv:2510.18917, Fig. 2

Clean vs. Reverberant Speech

Spectrogram comparison of three utterances — clean (left) vs. the same speech convolved with a room impulse response (right). The smearing of formant structure under reverberation is clearly visible.

arXiv:2601.19949, Fig. 5

Duration vs. RT60 Coverage

Utterance duration against RT60 across the reverberant corpus — long and short utterances span all RT60 bins, reducing confounding in word-error-rate analysis.

arXiv:2601.19949, Fig. 4

BeepBank-500 — Log-Mel Spectrograms

Example log-mel spectrograms across synthetic waveform families (sine, square, triangle, FM) and amplitude-modulation settings — the controlled parameter grid behind the earcon dataset.

arXiv:2509.17277, Fig. 1

Demos & Spaces

Live, interactive tools and leaderboards you can try in the browser.

Experience

A decade spanning automotive NVH, aeroacoustics, and large-scale industrial acoustics & audio AI.

Education, Skills & Honors

Let's build acoustic intelligence together

I enjoy connecting with researchers, engineers, and builders working on acoustics, audio ML, spatial audio, or machine perception.